PP-LinkNet: Improving Semantic Segmentation of High Resolution Satellite Imagery with Multi-stage Training

Abstract

Road network and building footprint extraction is essential for many applications such as updating maps, traffic regulations, city planning, ride-hailing, disaster response etc. Mapping road networks is currently both expensive and labor-intensive. Recently, improvements in image segmentation through the application of deep neural networks has shown promising results in extracting road segments from large scale, high resolution satellite imagery. However, significant challenges remain due to lack of enough labeled training data needed to build models for industry grade applications. In this paper, we propose a two-stage transfer learning technique to improve robustness of semantic segmentation for satellite images that leverages noisy pseudo ground truth masks obtained automatically (without human labor) from crowd-sourced OpenStreetMap (OSM) data. We further propose Pyramid Pooling-LinkNet (PP-LinkNet), an improved deep neural network for segmentation that uses focal loss, poly learning rate, and context module. We demonstrate the strengths of our approach through evaluations done on three popular datasets over two tasks, namely, road extraction and building footprint detection. Specifically, we obtain 78.19% meanIoU on SpaceNet building footprint dataset, 67.03% and 77.11% on the road topology metric on SpaceNet and DeepGlobe road extraction dataset, respectively. CCS CONCEPTS • Computing methodologies → Computer vision; Neural networks; • Information systems → Data mining.

Artificial Intelligence for Digital Heritage Innovation: Setting up a R&D Agenda for Europe

“…Another approach to automatically generating 3D/4D models comprises building footprint recognition and parametric modeling. Footprint recognition via semantic segmentation for aerial/satellite imagery [122][123][124][125] or from current cadastral data [126] and for contemporary photography [127] has been frequently researched. One issue in boundary detection workflows is overlapping building boundaries and texts.…”

A Digital 4D Information System on the World Scale: Research Challenges, Approaches, and Preliminary Results

“…Another approach to automatically generating 3D/4D models comprises building footprint recognition and parametric modelling. Footprint recognition via semantic segmentation for aerial/satellite imagery [150][151][152][153] or from current cadastral data [154] and for contemporary photography [155] has been frequently researched. One issue in boundary detection workflows is overlapping building boundaries and texts.…”

CSU-Net: Contour Semantic Segmentation Self-Enhancement for Human Head Detection

“…Wu et al [40] proposed extra learning of dilated affinity information in the DeepLab v3+ training to help the learning process and to refine it with a fast affinity propagation postprocessing, which exploits the extra information generated by the network. As well, Tran et al [41], developed a system using focal loss, poly learning rate, and context module to improve the robustness of semantic segmentation for satellite images. Finally, Gritzner and Ostermann [42] The survey of the state-of-the-art methods shows that in this field there are different deep learning approaches applied to improve the outcome of semantic segmentation.…”